Methods for detection of powdery mildew in agricultural plants with hyperspectral imaging

Abstract

With the human population still on the rise, it is important to continue developing technology for agriculture that will allow us produce enough food for the entire population. An important part of this is to become able to perform early detection of pathogens, as this will both reduce crop losses and reduce necessary amount of pesticide. This study was conducted to investigate methods involving hyperspectral image analysis for early detection and discrimination of powdery mildew on three greenhouse plants; cucumber, strawberry and tomato. Two separate experiments were conducted, one with a VNIR hyperspectral camera (400-1000 nm) and one with a SWIR hyperspectral camera (1000-2500 nm). For each plant type, multiple healthy and infected samples were scanned and analysed. The study had difficulty achieving a conclusive detection of powdery mildew with previously developed spectral disease indices. Results might improve if more appropriate methods for pre-processing and data filtering are applied. Principal component analysis (PCA) was performed in an attempt to identify informative spectral bands, and a PCA including a 1st derivative 2nd order polynomial Savitzky-Golay algorithm showed that wavelengths around 700-730 nm contributed the bulk of loadings, and thus >90% of the variation in an image with healthy and mildew-infected areas. This Parallel to the analysis performed in MATLAB, the graphical hyperspectral software Scyven was used as a supporting tool, and some comparisons were made to evaluate its usefulness as a tool for high precision scientific experiments. Though this proved useful for a beginner in the field of hyperspectral image analysis, its limitations makes it unsuitable for analysing the large image files used in this study.